Data driven approach combined with Hanasaki lake formulation

[PrateekSharma99]

While going through the mizuRoute code, I encountered the following statement:
case(2); ierr=20; message=trim(message)//'both data driven (follow target volume) and Hanasaki lake formulation are activated for a lake'; return

This error appears to indicate that the model does not permit the simultaneous activation of both Hanasaki lake formulation and target volume control for the same reservoir segment.

Based on the documentation and relevant literature, particularly from Gharari et al. (2024), it is stated that:

"For reservoirs in the river-lake network, users can specify target volumes corresponding to the model's simulation timestep (e.g., hourly or daily). The mizuRoute-Lake component then adjusts the outflow accordingly. If the specified target volume at a given timestep exceeds the current reservoir volume, the model will store water without releasing downstream. Conversely, if the reservoir volume surpasses the target volume, the model discharges the excess water." This statement raises a potential conflict when combining both Hanasaki volume computation and explicit target volume control.

Potential Issue:
If I use the Hanasaki lake formulation to simulate a reservoir for a specific date (e.g., 1st January 2022), while also specifying a control volume for the same date (e.g., 1st January 2022), will this dual activation result in the error mentioned above? This error seems to suggest that mizuRoute does not allow both methods to be active simultaneously for a single reservoir segment, as they would impose conflicting volume regulation mechanisms.

Clarification Sought:
What is the recommended approach to integrate both Hanasaki formulation and target volume control when both cannot be activated concurrently? Is there a methodology to toggle between the two approaches dynamically based on specific conditions or simulation periods?

[ShervanGharari]

@PrateekSharma99 this is a very good question!

For mizuRoute, in the current implementation, you have to choose between:

• Following a target volume (it should be provided at model time step, as stated in the paper), or
• Using one of the built-in formulations.

---

Suggested Approach

For what you're trying to achieve, I believe you should use the restart capability of the model:

1. Run the model up to 31 December 2021 and output the restart file.
2. Manipulate the volume of the lakes in that restart file.
3. Use it as the restart input for a new simulation beginning 1 January 2022.

This way, you can:

• Use Hanasaki (or any formulation),
• Adjust lake storage at specific times and locations fully under your control.

---

Restart File Settings

To output a restart file at the end of the simulation (or at periodic timestamps like monthly, yearly, etc.), set the control flag:

• restrat_write

To restart a simulation using a restart file, set the flag:

• fname_state_in

---

There may be users or developers working on data assimilation.
Perhaps @nmizukami can provide more insight, including any related publications or repositories, or suggest other approaches.

[PrateekSharma99]

@ShervanGharari Thank you for the clarification regarding the mutual exclusivity of Hanasaki formulation and target volume control in mizuRoute.

Building on that—what would be the recommended approach to handle a case where I have observed volume time series for N reservoirs, but the data is incomplete or intermittent (i.e., with missing periods)?

Specifically:

Is there a way to run a long continuous simulation (e.g., multi-decadal), where target volume control is only applied at timesteps where observations are available, and the model falls back to a default (e.g., Hanasaki or another built-in formulation) when observations are missing?

Would this still require restarting the model at each observation point, or is there a way to handle such partial target volume inputs without repeatedly breaking the simulation?

The goal is to keep the setup scalable and efficient across many reservoirs with variable observation availability.

Thanks in advance for your guidance!

[ShervanGharari]

I think the most reasonable approach is to pause and resume the simulation while modifying the mizuRoute restart file. For example, if your domain includes 22 lakes, you could update the 5 for which you have data on January 1st, 2022. Then, on January 17th, you might update 2 additional reservoirs, and so on. This process is manageable by leveraging daily restart outputs from mizuRoute—simply update the restart file at the desired times for the desired reservoir and continue the simulation. A Python wrapper could efficiently automate this workflow.
That said, I haven’t personally used mizuRoute for data assimilation, so others may have more robust or straightforward solutions. Additionally, I’m not certain how this approach would interact with the Hanasaki formulation, in case it introduces any unexpected behavior.
Hope it helps.

[nmizukami]

If I read this correctly, I understand your intention is to use Hanasaki to simulate the lakes, and use observed volume at some lakes and some time steps available to ingest the observation into the simulation.

There is some capability in mizuRoute of doing direct-insertion of observed discharge at river reach whenever and wherever observation is available (PR 354 for some test), but not for volume now. And, it has not been tested for real application (this was request from Andy W for Columbia River forecast).

We could change the codes to enable target volume and hanasaki at the same lake (may take some time to implement and test); similar to discharge direct insertion DA, when and where you have observed lake volume, use obs as target volume and not running hanasaki (I guess for this it makes sense to have some longer time period of obs, not just one time step), but use the volume read from file to guide to control the discharge and volume. target volume may not bring your volume to the volume you want immediately. If you want to bring the simulated volume to obs volume, this would be regular data assimilation problem (so direct-insertion is the easiest way). Also, either way occasional target volume or DA, not sure how the simulation looks like. For DA, you might see abrupt change in simulated time series. If your goal of the simulation is something like forecast, DA makes sense, but if it is to generate some retrospective continuous simulations, may not look good.

As Shervan suggested, you could test using restart with modified volume (this is the same effect as direct insertion), and see what happen.

[PrateekSharma99]

Good morning @nmizukami I was going through literature and could not find any DA for reservoir storage that has been done in mizuRoute as of now. In theory, this seems doable. But is there a downside to it that I am missing?

[nmizukami]

Hi Prateek, a direct-insertion DA for river reach exists in mizuRoute, but no publication for this. For DA for reservoir storage, there has been no work toward this. There is no downside I am aware of, but just do not have a funding to do this specific thing.

At what stage is your work now? and are you still trying to figure out the way to do DA for reservoir storage?

[PrateekSharma99]

Good Morning @nmizukami, currently my reservoir integrated model is set up and working. Now I am finally figuring my way forward for DA in reservoir storage, which is my goal. Since specifically for mizuroute, there is no literature for storage DA is there any suggestions for this work because I am currently figuring it out myself.